Brain Tumor Detection Eval

brain-tumor-detection-eval

Deep Brain Net: An Optimized Deep Learning Model for Brain tumor Detection in MRI Images Using EfficientNetB0 and ResNet50 with Transfer Learning — Onah et al. (2025) (arXiv:2507.07011, 2025)

What this evaluates

Evaluates a deep learning model's ability to classify brain MRI images into four tumor categories (glioma, meningioma, no tumor, pituitary). It probes multi-class image classification performance, generalization to unseen medical scans, and the model's capacity to balance precision and recall across classes.

Datasets

• Public MRI dataset (unspecified) — total ?; splits: train (-1), val (-1)

Metrics

• accuracy (primary) — range: [0, 1]
  • Proportion of correctly classified instances out of the total number of instances.
• weighted F1-score — range: [0, 1]
  • Harmonic mean of precision and recall, averaged across classes weighted by the number of true instances per class.
• macro-average AUC-ROC — range: [0, 1]
  • Area under the Receiver Operating Characteristic curve computed for each class using a one-vs-rest approach, then averaged unweighted across all classes.

Input / output format

Input: Grayscale MRI images converted to RGB via channel stacking, resized to 224×224 pixels, and standardized using EfficientNet's normalization scheme.

Output: Predicted tumor class label from four categories: Glioma tumor, Meningioma tumor, No tumor, Pituitary tumor.

Scoring recipe

def compute_metrics(y_true, y_pred, y_prob, num_classes=4):
    accuracy = (y_true == y_pred).mean()
    precision, recall, f1 = precision_recall_fscore_support(y_true, y_pred, average=None)
    weighted_f1 = f1_score(y_true, y_pred, average='weighted')
    auc_roc = roc_auc_score(y_true, y_prob, multi_class='ovr', average='macro')
    return {'accuracy': accuracy, 'weighted_f1': weighted_f1, 'macro_auc_roc': auc_roc}

Common pitfalls

• The paper reports conflicting validation accuracy values (94.7% vs 93.2%) without clarifying which corresponds to the final classification report.
• No explicit train/validation/test split sizes or ratios are provided, making reproducibility of data partitioning difficult.
• Grayscale images are converted to 3-channel RGB for pre-trained backbones, which may introduce redundant channel information not accounted for in standard ImageNet pre-training protocols.

Evidence (verbatim from paper)

The classification report, shown in Table III below, summarizes the model's performance in each class using three key metrics: precision, recall and F1 score. ... The weighted F1-score across all classes, weighted by the number of true instances per class, is 0.88. ... Additionally, the AUC-ROC score provides insight into the model's ability to distinguish between classes across different threshold values, offering a comprehensive measure of classification performance. The macro-average AUC-ROC score is 0.98...

Citation

@misc{onah2025deepbrainnet,
  title={Deep Brain Net: An Optimized Deep Learning Model for Brain tumor Detection in MRI Images Using EfficientNetB0 and ResNet50 with Transfer Learning},
  author={Onah et al. (2025)},
  year={2025},
  note={arXiv:2507.07011}
}

• arXiv: 2507.07011